Musk Calls Claude Opus 5.5 a 'Masterpiece', Admits Grok 4.8 Falls Short

Elon Musk called Anthropic's Claude Opus 5.5 a "masterpiece" and admitted his own company's next model won't beat it, in a rare public concession from a competitor. "Opus 5.5 is a masterpiece," Musk wrote on X, adding that Grok 4.8 would be "not as good", according to Decrypt.

He hedged on the longer term, predicting that Grok 5 "should exceed Opus 5.5" within "two or three months" — a target with no firm date attached. Anthropic released Opus 5.5 on September 22, and the company has not publicly responded to Musk's comments.

The admission comes with a revealing technical detail: xAI's team abandoned JAX and SGLang and rewrote its inference and reinforcement‑learning system in C and C++. Musk described the language choice as "C with a little C++," crediting Starlink engineers for work on the training stack, including custom headers and precise time calculations for leap seconds.

Grok 4.8's pretraining is finished but it's still in reinforcement learning. Musk called the current state "not yet incredible" but said the foundations are "built right." A soft rollout to X Bot users was expected later this week, Decrypt reported, and no benchmark scores have been published for Grok 4.8.

The numbers back up Musk's humility. On Terminal‑Bench 4.0, which measures command‑line agent tasks, Claude Opus 5.5 scored 66.4% (Anthropic's self‑reported figure) — well ahead of OpenAI's GPT‑6 Astra at 57.9% — while Grok trails far behind with no published scores, according to Decrypt's reporting.

Opus 5.5 also runs 40% cheaper than the previous Opus while matching the higher‑end Fable 5.1 on most tasks. For builders deciding where to spend their API budget, that combination — stronger agent performance at a lower price — is exactly the kind of shift that moves workloads.

The real signal isn't Musk's post — it's the spend data. Claude Opus 5.5 was the top model by paid spend in its first full month, capturing roughly 14% of model spend in October, according to the Vercel AI Gateway Production Index cited by Yahoo Tech. Anthropic's models dominate paid usage: Sonnet 5.5 led at 26% of teams, followed by Haiku 5.5 at 22%, Haiku 4.5 at 20%, and Opus 5.5 at 18%.

Vercel also found that AI models now have a half‑life of about six weeks — teams adopt a new model and switch to the next release within roughly a month. That churn is why xAI's multi‑month delay on Grok 4.8 stings: by the time it ships, the builder community may have already moved on.

Musk publicly ranking a rival's model above his own is more than a compliment — it's an admission that xAI has fallen behind on the agentic coding front that now drives enterprise adoption. Anthropic's $17 billion in Q3 revenue, against a $567 million operating loss, shows the company has found a path to efficiency that OpenAI's $10.9 billion loss can't match. xAI's full‑stack rewrite suggests Musk is betting that raw systems engineering — not just more training compute — is the way back.

The Vercel data cuts against the narrative that this is a two‑horse race between OpenAI and Anthropic. Grok isn't showing up in the top paid‑model lists, which means for most builders the practical choice right now is Claude versus GPT — with Opus 5.5 currently winning on the agentic tasks that matter most.